<![CDATA[One of the primary factors impacting the economy is the ability to forecast the prices of commodities such as beef. This paper aims to evaluate the effectiveness of various activation functions and optimization strategies when integrated into the LSTM (Long Short-Term Memory) architecture model in predicting the price of lean beef in Aceh. The data sample utilized was obtained from the Indonesian National Food Agency panel, which shows daily prices for beef within the time frame of July 14th, 2022, to July 31st, 2024. As for the conducted research, the process of preparation data preprocessing, partitioning data into training, validation and test sets and the actual execution of the LSTM model which was trained using four different types of activation functions: tanh, ReLU, sigmoid and PReLU together with three different optimizers: Adam, Nadam and RMSprop for 50, 70, 100 and 200 training iterations. The evaluation metrics employed were Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R-squared). The best performance was recorded at 200 epochs with the combination of PReLU activation function and Nadam optimizer, which had the best performance with RMSE = 2.56, MAPE = 0.65% and R² = 0.104. This combination was more effective than others since it depicted better overall performance in identifying complex non-linear relationships that existed in the price data. Further on, Nadam seems to have benefits in terms of allowing the model to converge faster and making the training more stable. This work stresses the selection of activation functions and optimization methods when building LSTM models aimed at forecasting prices of commodities with large volatility. It will be very helpful in developing better predictive models and decision-making processes in the agro-business. Another way to enhance predictive performance could be changing the model architecture or using different techniques, such as attention mechanisms.]]>
Copyrights © 2025
